By Wentao Huang
InfoNets is a Pytorch-based Python package that provides a high-level neural networks API and the efficiency algorithms for deep learning which is not based on the traditional BP (Back Propagation) algorithm.
TensorFlow, Pytorch and other opensource deep learning software frameworks provide great convenience for current deep learning researchers and developers, and make it easy to study and develop deep learning related algorithms and products. However, they only provide good encapsulation and support for traditional deep learning that is based on BP algorithm, and it is not very convenient for non-BP-based deep learning method.
Based on this, I developed this framework and software package, which can well support the non-BP-based deep learning, and also integrate the BP algorithm into the same framework at the same time. Moreover, it provides the more powerful support for the operation and management of general data sets.
On the other hand, this framework is highly integrated into the non-BP-based deep learning algorithms developed by me, which is based on information theory and brain-inspired computing and can train a deep neural network quickly from the bottom-up and layer-by-layer. It can simultaneously process a large number of training samples in parallel, which is very suitable for GPU operation and can greatly improve the efficiency and effectiveness for training deep networks. Furthermore, it is very easy to make secondary development based on this package.